Research map / Learning theory and algorithms / Prediction and decision losses

Performative Risk Control: Calibrating Models for Reliable Deployment under Performativity

NeurIPS 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Probabilistic calibration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Performative Risk Control (PRC), an iterative threshold-calibration procedure that provides finite-sample risk control under performative distribution shift, extended to quantile risks (CVaR), and demonstrated on simulated credit scoring.

Problem

Risk control of black-box model predictions with finite-sample guarantees is studied only under static distributions, but predictions can change the outcome distribution (performativity), so guarantees fail under strategic responses.

Approach

An iteratively refined threshold calibration that, at each round, deploys the current threshold, collects losses from the induced distribution, and lowers the threshold only as far as a UCB-based risk bound plus a performativity guard (tau times threshold change, with tau >= gamma, the Lipschitz sensitivity of the distribution map) allows, so every deployed threshold is risk-controlled.

Key results

• On a class-balanced credit scoring subset with n=2000, alpha=0.3, delta=0.1, PRC iterates safely and converges to a risk-controlled, non-conservative threshold (empirically gamma <= 1.38).
• For 90%-CVaR type II error control with n=10,000, alpha=0.25, PRC ends with final risk inside the confidence bounds; empirically gamma <= 0.205, so tau >= 2.05 is required by the theory.
• The asymptotic tightness gap Delta-alpha shrinks as O(ln n / sqrt(n)).

Surprising findings

• The simple post-processing may output lambda_safe (no useful threshold), since calibration only postprocesses and cannot fine-tune the model.

Limitations

• Guarantees rely on a known or estimated Lipschitz sensitivity gamma; tau must be chosen to exceed it, which is hard to know a priori.
• Simple threshold post-processing may be less effective than fine-tuning the whole model and can return lambda_safe.
• Evaluation uses a simulated performative shift on a single credit scoring dataset.

Experimental setup

Datasets: Kaggle credit scoring dataset

Benchmarks: Kaggle credit scoring (simulated performative shift)

Metrics: performative expected risk type II error 90%-CVaR

Builds on: conformal risk control performative prediction Hoeffding-Bentkus inequality

Closest papers by meaning

• Statistical Inference under Performativity (NeurIPS 2025 · similarity 0.87)
• On the Impact of Performative Risk Minimization for Binary Random Variables (ICML 2025 · similarity 0.86)
• Distributionally Robust Performative Prediction (NeurIPS 2024 · similarity 0.85)
• Plug-in Performative Optimization (ICML 2024 · similarity 0.85)
• Tight Lower Bounds and Improved Convergence in Performative Prediction (NeurIPS 2025 · similarity 0.85)
• Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.85)